Project Page | Arxiv - CVPR 2023
Motion Latent Diffusion (MLD) is a text-to-motion and action-to-motion diffusion model. Our work achieves state-of-the-art motion quality and two orders of magnitude faster than previous diffusion models on raw motion data.
- [2023/06/20] MotionGPT is released! A unified motion-language model. Do all your motion tasks in MotionGPT
- [2023/03/08] add the script for latent space visualization and the script for the floating point operations (FLOPs)
- [2023/02/28] MLD got accepted by CVPR 2023!
- [2023/02/02] release action-to-motion task, please refer to the config and the pre-train model
- [2023/01/18] add a detailed readme of the configuration
- [2023/01/09] release no VAE config and pre-train model, you can use MLD framework to train diffusion on raw motion like MDM.
- [2022/12/22] first release, demo, and training for text-to-motion
- [2022/12/08] upload paper and init project, code will be released in two weeks
Setup and download
conda create python=3.9 --name mld
conda activate mld
Install the packages in requirements.txt and install PyTorch 1.12.1
pip install -r requirements.txt
We test our code on Python 3.9.12 and PyTorch 1.12.1.
Run the script to download dependencies materials:
bash prepare/download_smpl_model.sh
bash prepare/prepare_clip.sh
For Text to Motion Evaluation
bash prepare/download_t2m_evaluators.sh
Run the script to download the pre-train model
bash prepare/download_pretrained_models.sh
Visit the Google Driver to download the previous dependencies and model.
Text-to-motion
We support text file or keyboard input, the generated motions are npy files.
Please check the configs/asset.yaml for path config, TEST.FOLDER as output folder.
Then, run the following script:
python demo.py --cfg ./configs/config_mld_humanml3d.yaml --cfg_assets ./configs/assets.yaml --example ./demo/example.txt
Some parameters:
--example=./demo/example.txt: input file as text prompts--task=text_motion: generate from the test set of dataset--task=random_sampling: random motion sampling from noise--replication: generate motions for same input texts multiple times--allinone: store all generated motions in a single npy file with the shape of[num_samples, num_ replication, num_frames, num_joints, xyz]
The outputs:
npy file: the generated motions with the shape of (nframe, 22, 3)text file: the input text prompt
Training guidance
Please refer to HumanML3D for text-to-motion dataset setup. We will provide instructions for other datasets soon.
Please first check the parameters in configs/config_vae_humanml3d.yaml, e.g. NAME,DEBUG.
Then, run the following command:
python -m train --cfg configs/config_vae_humanml3d.yaml --cfg_assets configs/assets.yaml --batch_size 64 --nodebug
Please update the parameters in configs/config_mld_humanml3d.yaml, e.g. NAME,DEBUG,PRETRAINED_VAE (change to your latest ckpt model path in previous step)
Then, run the following command:
python -m train --cfg configs/config_mld_humanml3d.yaml --cfg_assets configs/assets.yaml --batch_size 64 --nodebug
Please first put the tained model checkpoint path to TEST.CHECKPOINT in configs/config_mld_humanml3d.yaml.
Then, run the following command:
python -m test --cfg configs/config_mld_humanml3d.yaml --cfg_assets configs/assets.yaml
Render SMPL
Refer to TEMOS-Rendering motions for blender setup, then install the following dependencies.
YOUR_BLENDER_PYTHON_PATH/python -m pip install -r prepare/requirements_render.txt
Run the following command using blender:
YOUR_BLENDER_PATH/blender --background --python render.py -- --cfg=./configs/render.yaml --dir=YOUR_NPY_FOLDER --mode=video --joint_type=HumanML3D
python -m fit --dir YOUR_NPY_FOLDER --save_folder TEMP_PLY_FOLDER --cuda
This outputs:
mesh npy file: the generate SMPL vertices with the shape of (nframe, 6893, 3)ply files: the ply mesh file for blender or meshlab
Run the following command to render SMPL using blender:
YOUR_BLENDER_PATH/blender --background --python render.py -- --cfg=./configs/render.yaml --dir=YOUR_NPY_FOLDER --mode=video --joint_type=HumanML3D
optional parameters:
--mode=video: render mp4 video--mode=sequence: render the whole motion in a png image.
Solve foot sliding issue
If your demo results have a severe issue on foot sliding, please take a look to the below. It could happen when self.feats2joints (use mean and std for de-normalization) is broken.
motion-latent-diffusion/mld/data/get_data.py
Lines 26 to 41 in 5c264c3
Details of training
- GPUs. You can indicate the IDs to use all your GPUs.
- Epoch Nums. 1500~3000 epoch is enough for VAE or MLD. I suggest you use wandb(prefer) or tensorborad to check FID curve of your training.
- Training Speed. 2000 epoch could cost 1 day for a single GPU, and around 12 hours for 8 GPUs. Training speed also depends on
VAL_EVERY_STEPS(Validation Frequency), DataIO Speed. Your training is a little slow. - Data Log. Only loss print by default. After validation, more metrics of val will print. More details in wandb (prefer) or tensorborad.
- Debug or not. Please use
--nodebugfor all your training. - VAE loading. Please load your pre-train VAE correctly for the MLD diffusion training.
- FID. FID of validation will drop to 0.5~1 after 1500 epochs for both VAE and MLD training. By default, validation is on test split...
Details of motion lengths
Our model is capable of generating motions with arbitrary lengths. To handle different lengths of motions in the same batch, padding and masking are utilized in our motion encoder and decoder. After latent vector z is obtained by diffusion process, motion length L represented as a sequence of positional encodings in the form of sinusoidal functions are also provided to the motion decoder, so our motion decoder is able to generate output with variable target lengths.MLD-1 VS MLD-7
MLD-7 only works best in evaluating VAE models (Tab. 4), and MLD-1 wins these generation tasks (Tab. 1, 2, 3, 6). In other words, MLD-7 wins the first training stage for the VAE part, while MLD-1 wins the second for the diffusion part. We thought MLD-7 should perform better than MLD-1 in several tasks, but the results differ. The main reason for this downgrade of a larger latent size, we believe, is the small amount of training data. HumanML3D only includes 15k motion sequences, much smaller than billions of images in image generation. MLD-7 could work better when the motion data amount reaches the million level.Details of Inference Time
We provide a detailed ablation study with DDIM below. We evaluate the total inference time to generate 2048 motion clips with different diffusion schedules, floating point operations (FLOPs) counted by THOP library, the size of diffusion input, and FID. MLD reduces the computational cost of diffusion models, which is the main reason for faster inference. The iterations of diffusion further widen the gap in computational cost.python -m scripts.flops --cfg configs/your_config.yaml
Latent Space Visualization
We provide Visualization of the t-SNE results on evolved latent codes zt during the reverse diffusion process (inference) on action-to-motion task below. t is the diffusion step but ordered in the forward diffusion trajectory. zt=49 is the initial random noise. zt=0 is our prediction. We sample 30 motions for each action label. From left to right, it shows the evolved latent codes during the inference of diffusion models.python -m scripts.tsne --cfg configs/your_config.yaml
Note: This only support action-to-motion models for now.
If you find our code or paper helps, please consider citing:
@inproceedings{chen2023executing,
title={Executing your Commands via Motion Diffusion in Latent Space},
author={Chen, Xin and Jiang, Biao and Liu, Wen and Huang, Zilong and Fu, Bin and Chen, Tao and Yu, Gang},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={18000--18010},
year={2023}
}Thanks to TEMOS, ACTOR, HumanML3D and joints2smpl, our code is partially borrowing from them.
This code is distributed under an MIT LICENSE.
Note that our code depends on other libraries, including SMPL, SMPL-X, PyTorch3D, and uses datasets which each have their own respective licenses that must also be followed.